← All problems

Dimension-free consistency of empirical spatial depth

Less than 1 year old · traced to

Theorem 2.1. For a sample of iid random elements X1,…,Xn∼PXX_{1},\ldots,X_{n}\sim P^{X} with values in Rd\mathbb{R}^{d}, the empirical spatial distribution estimator and the plug-in spatial-depth estimator satisfy

E(sup⁡x∈Rd∣DS(x,P^nX)−DS(x,PX)∣)≤E(sup⁡x∈Rd∥S^x−Sx∥Rd)≤4πn.\mathbb{E}\left(\sup_{x\in\mathbb{R}^{d}}\left|D_{S}(x,\widehat{P}_{n}^{X})-D_{S}(x,P^{X})\right|\right)\leq\mathbb{E}\left(\sup_{x\in\mathbb{R}^{d}}\left\|\widehat{S}_{x}-S_{x}\right\|_{\mathbb{R}^{d}}\right)\leq\frac{4\sqrt{\pi}}{\sqrt{n}}.

Thus the expected uniform error is bounded by a rate depending only on nn, independently of the ambient dimension dd and without boundedness, moment, smoothness, tuning-parameter, or regularization assumptions, as stated in the source paper.

References

Progress summary

Refreshed
Open · solution reported

A 2026 preprint proves the desired consistency bound with a rate that does not worsen as the dimension grows.

The problem asks whether empirical spatial distribution and plug-in spatial-depth estimators converge uniformly in L1L^1 at a dimension-independent rate. A 2026 arXiv preprint gives an affirmative theorem for i.i.d. data in Rd\mathbb{R}^{d}, and extends the argument to separable Hilbert spaces.

Known results

  • The spatial distribution function originated with Chaudhuri (1996).
  • Koltchinskii (1997, Theorem 2.5) proved that it uniquely characterizes the underlying probability measure.
  • Spatial-depth foundations include Vardi and Zhang (2000), Serfling (2002), and Gao (2003).
  • González-Sanz and Konen (2026) proved that spatial depth itself uniquely characterizes the underlying probability measure.

2026 dimension-free consistency theorem

The preprint proves E[sup⁡x∈Rd∣DS(x,P^nX)−DS(x,PX)∣]≤4π/n\mathbb{E}[\sup_{x\in\mathbb{R}^{d}}|D_S(x,\widehat{P}_n^X)-D_S(x,P^X)|]\leq 4\sqrt{\pi}/\sqrt{n}, using a dimension-independent Rademacher bound. No boundedness, moment, smoothness, covering-number, packing-number, or compactness assumptions are required.

Current status (as of July 2026): The question is settled affirmatively by the arXiv proof, with rate 4π/n4\sqrt{\pi}/\sqrt{n} independent of dd; no remaining mathematical gap is reported.

Sources

Solutions 0

No solutions have been posted yet.